Looper Insight 9

Demo and values

One MVIS camera + IMU run on hand-held recording 1: the two mono cameras, the rolling-shutter colour camera and its readout time, solved together.

The scene MVIS recovered from this dataset, as on the homepage. The values beside it come from this page's run.

Reprojection median0.17 px

cam0–cam1 baseline99.2 mm

cam0 td+1.24 ms

cam0 lever arm43.9 mm

Reprojection is how far, in pixels, the detected corners lie from where the solve puts them. td is the camera–IMU time offset, and the lever arm is the camera–IMU position offset.

Calibration process

MVIS's own player, from this run: the detected board corners, then the rig's path as the solve refines it. Press Play or drag the bar; pick a camera card to look through that camera. We publish no camera images.

Solved rig

The solved positions and orientations, to scale in the IMU frame. Drag sideways to orbit. Tap a sensor to see its values.

3DFrontTopSide+−

Sensor	fx, fy, cx, cy px	Lever mm	log(R) rad	td ms	Readout ms	t_ga ms	Reprojection px
cam0	334.1, 334.1, 272.0, 318.5	[−12.5, +38.4, +17.3]	[+0.0109, +0.0025, −1.5656]	+1.24	global		0.161
cam1	334.0, 333.9, 269.6, 318.4	[−12.2, −60.8, +16.7]	[+0.0080, +0.0023, −1.5642]	+1.25	global		0.167
cam2	775.1, 775.2, 960.1, 535.7	[−12.5, −11.2, +16.1]	[+0.0066, +0.0108, −3.1361]	−5.84	32.36		0.449
IMU (base)		0	0	0		−5.08 ± 0.04

Lever arms and rotations are each sensor in the base IMU frame; rotation is log(R) in rad. td is the time offset to the base IMU, t_imu = t_sensor + td. Reprojection is the median per camera (all cameras: median 0.175 px): how well the model fits the corners, not how accurate it is.

Check against a reference

No published reference. No published calibration exists for this device, so this page has nothing to compare against. What we can show is how much each value moves over the 10 hand-held recordings. All 10 recordings and tools: Dataset eval →

Config and run

looper_insight9_bag1_cam_imu.yaml

input:
  format: euroc
  path: 1/mav0
  cam_topics:
  - /cam0
  - /cam1
  - /cam2
  imu_topics:
  - /imu0
  window_mode: manual
  t0: 0.0
  window: 0.0
board:
  board_type: apriltag
  rows: 6
  cols: 6
  tag_size: 0.055
  tag_spacing: 0.3
  tag_family: t36h11
camera_input:
  path: insight9_seed_calib.yaml
  free_intrinsics: true
  estimate_readout: false
  cameras:
  - {}
  - {}
  - estimate_readout: true
    rolling_mode: 0
solve:
  state_hz: 30.0
  max_iterations: 40
  imu_intrinsics: dw_da
  imu_intrinsics_model: calyx
  corner_subsample: top_left
  estimate_t_ga: true
  sigma_t_ga: 0.01
  init_kwargs:
    keyframe_hz: 2.0
    scale_rel_tol: 0.3
  sigma_g: 0.00021
  sigma_a: 0.0032
  sigma_bg: 0.001
  sigma_ba: 0.002
output:
  path: cam_imu.json

shell

mvis cam-imu -c looper_insight9_bag1_cam_imu.yaml

The marked lines are the choices that matter for this rig:

• camera_input.free_intrinsics: Camera intrinsics stay free in the camera + IMU solve.
• camera_input.cameras: The colour camera (cam2) has a rolling shutter, so its readout time is estimated.
• solve.estimate_t_ga: Estimates t_ga, the time offset between the gyroscope and the accelerometer.
• solve.sigma_g: IMU noise from a still recording of this unit, measured density × 2.5. The other three sigmas are set the same way.

This is the config of the run on this page, with every path cut to its file name. Point the paths at your copy.

The solve took about 6 min on one desktop. It writes cam_imu.json, a calibration YAML and a report beside the config.

MVIS

MVIS main, measured-noise batch of 2026-11-08

Run date

2026-11-08

Data

Calibench, looper-insight9/ →

Config

looper_insight9_bag1_cam_imu.yaml, above

Command

mvis cam-imu -c looper_insight9_bag1_cam_imu.yaml

Dataset

• Hand-held recording 1 of 10, about 66 s. The run uses the whole recording.
• Cameras: two 544 × 640 global-shutter mono cameras at 20 Hz and a 1920 × 1088 rolling-shutter colour camera at about 4.3 Hz, all fitted with the equidistant lens model. IMU at 400 Hz.
• Board: AprilGrid 6×6, tag 55 mm, spacing 0.3, t36h11.
• Access is by automatic approval on the Hub; the data is CC BY-NC 4.0, commercial use by request.
• Solve window: 66.4 s.

Download: Calibench, looper-insight9/ → The recordings are listed there, not copied here.

AprilGrid 6 × 6, 412 × 412 mm. Standard pattern, to scale

More on this data

Chessboard and robot-arm recordings Device wiki

The sensor page also shows our own unit on a chessboard and two robot-arm recordings, each solved on its own. Looper Insight 9 →